Re: [PATCH tip/core/rcu 16/27] rcu: Add comment documenting how rcu_seq_snap works

[Joel Fernandes]

On Tue, Jun 26, 2018 at 07:30:55PM +0200, Peter Zijlstra wrote:
> On Mon, Jun 25, 2018 at 05:35:02PM -0700, Paul E. McKenney wrote:
> > +/*
> > + * rcu_seq_snap - Take a snapshot of the update side's sequence number.
> > + *
> > + * This function returns the earliest value of the grace-period sequence number
> > + * that will indicate that a full grace period has elapsed since the current
> > + * time.  Once the grace-period sequence number has reached this value, it will
> > + * be safe to invoke all callbacks that have been registered prior to the
> > + * current time. This value is the current grace-period number plus two to the
> > + * power of the number of low-order bits reserved for state, then rounded up to
> > + * the next value in which the state bits are all zero.
> 
> If you complete that by saying _why_ you need to round up there, then
> the below verbiage is completely redundant.
> 
> > + * In the current design, RCU_SEQ_STATE_MASK=3 and the least significant bit of
> > + * the seq is used to track if a GP is in progress or not. Given this, it is
> > + * sufficient if we add (6+1) and mask with ~3 to get the next GP. Let's see
> > + * why with an example:
> > + *
> > + * Say the current seq is 12 which is 0b1100 (GP is 3 and state bits are 0b00).
> > + * To get to the next GP number of 4, we have to add 0b100 to this (0x1 << 2)
> > + * to account for the shift due to 2 state bits. Now, if the current seq is
> > + * 13 (GP is 3 and state bits are 0b01), then it means the current grace period
> > + * is already in progress so the next GP that a future call back will be queued
> > + * to run at is GP+2 = 5, not 4. To account for the extra +1, we just overflow
> > + * the 2 lower bits by adding 0b11. In case the lower bit was set, the overflow
> > + * will cause the extra +1 to the GP, along with the usual +1 explained before.
> > + * This gives us GP+2. Finally we mask the lower to bits by ~0x3 in case the
> > + * overflow didn't occur. This masking is needed because in case RCU was idle
> > + * (no GP in progress so lower 2 bits are 0b00), then the overflow of the lower
> > + * 2 state bits wouldn't occur, so we mask to zero out those lower 2 bits.
> > + *
> > + * In other words, the next seq can be obtained by (0b11 + 0b100) & (~0b11)
> > + * which can be generalized to:
> > + * seq + (RCU_SEQ_STATE_MASK + (RCU_SEQ_STATE_MASK + 1)) & (~RCU_SEQ_STATE_MASK)
> > + */
> 
> Is the below not much simpler:
> 
> >  static inline unsigned long rcu_seq_snap(unsigned long *sp)
> >  {
> >  	unsigned long s;
> 
> 	s = smp_load_aquire(sp);
> 
> 	/* Add one GP */
> 	s += 1 << RCU_SEQ_CTR_SHIFT;
> 
> 	/* Complete any pending state by rounding up */

I would suggest this comment be changed to "Add another GP if there was a
pending state".

> 	s = __ALIGN_MASK(s, RCU_SEQ_STATE_MASK);
> 

I agree with Peter's suggestions for both the verbiage reduction in the
comments in the header, as the new code he is proposing is more
self-documenting. I believe I proposed a big comment just because the code
wasn't self-documenting or obvious previously so needed an explanation.

How would you like to proceed? Let me know what you guys decide, I am really
Ok with anything. If you guys agree, should I write a follow-up patch with
Peter's suggestion that applies on top of this one?  Or do we want to drop
this one in favor of Peter's suggestion?

I guess we also have to conclude the other part about using memory barriers,
but I think that should be a separate patch.

thanks!

 - Joel